Re: [PATCH] block: Improve shared tag set performance

[Bart Van Assche <bvanassche@acm.org>]

On 10/20/23 18:31, Ming Lei wrote:
> If two LUNs are attached to same host, one is slow, and another is fast,
> and the slow LUN can slow down the fast LUN easily without this fairness
> algorithm.
> 
> Your motivation is that "One of these logical units (WLUN) is used
> to submit control commands, e.g. START STOP UNIT. If any request is
> submitted to the WLUN, the queue depth is reduced from 31 to 15 or
> lower for data LUNs." I guess one simple fix is to not account queues
> of this non-IO LUN as active queues?

Hi Ming,

For fast storage devices (e.g. UFS) any time spent in an algorithm for
fair sharing will reduce IOPS. If there are big differences in the
request processing latency between different request queues then fair
sharing is beneficial. Whether or not the fair sharing algorithm is
improved, how about making it easy to disable fair sharing, e.g. with
something like the untested patch below? I think that will work better
than ignoring fair sharing per LUN. UFS devices support multiple logical
units and with the current fair sharing approach it takes long until
tags are taken away from an inactive LUN (request queue timeout).

Thanks,

Bart.


diff --git a/block/blk-mq.h b/block/blk-mq.h
index f75a9ecfebde..b06b161d06de 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -416,7 +416,8 @@ static inline bool hctx_may_queue(struct 
blk_mq_hw_ctx *hctx,
  {
  	unsigned int depth, users;

-	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
+	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
+	    hctx->queue->disable_fair_sharing)
  		return true;

  	/*
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index eef450f25982..63b04cf65887 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -523,6 +523,7 @@ struct request_queue {
  	struct mutex		debugfs_mutex;

  	bool			mq_sysfs_init_done;
+	bool			disable_fair_sharing;
  };

  /* Keep blk_queue_flag_name[] in sync with the definitions below */